1 /* Copyright (c) 2008-2015. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
14 #include "src/internal_config.h"
15 #include "src/mc/mc_private.h"
16 #include "xbt/module.h"
17 #include <xbt/mmalloc.h>
18 #include <xbt/memory.hpp>
19 #include "src/smpi/private.h"
21 #include "src/xbt/mmalloc/mmprivate.h"
23 #include "src/simix/smx_private.h"
25 #include <libunwind.h>
28 #include "src/mc/mc_private.h"
31 #include "src/mc/mc_snapshot.h"
32 #include "src/mc/mc_mmu.h"
33 #include "src/mc/mc_unw.h"
34 #include "src/mc/mc_protocol.h"
35 #include "src/mc/mc_smx.h"
36 #include "mc_hash.hpp"
38 #include "src/mc/RegionSnapshot.hpp"
39 #include "src/mc/ObjectInformation.hpp"
40 #include "src/mc/Frame.hpp"
41 #include "src/mc/Variable.hpp"
43 using simgrid::mc::remote;
45 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(mc_checkpoint, mc,
46 "Logging specific to mc_checkpoint");
51 /************************************ Free functions **************************************/
52 /*****************************************************************************************/
54 /** @brief Restore a region from a snapshot
56 * @param reg Target region
58 static void restore(mc_mem_region_t region)
60 switch(region->storage_type()) {
61 case simgrid::mc::StorageType::NoData:
63 xbt_die("Storage type not supported");
66 case simgrid::mc::StorageType::Flat:
67 mc_model_checker->process().write_bytes(region->flat_data().get(),
68 region->size(), region->permanent_address());
71 case simgrid::mc::StorageType::Chunked:
72 mc_region_restore_sparse(&mc_model_checker->process(), region);
75 case simgrid::mc::StorageType::Privatized:
76 for (auto& p : region->privatized_data())
83 RegionSnapshot privatized_region(
84 RegionType region_type, void *start_addr, void* permanent_addr,
85 std::size_t size, const RegionSnapshot* ref_region
88 size_t process_count = MC_smpi_process_count();
90 // Read smpi_privatisation_regions from MCed:
91 smpi_privatisation_region_t remote_smpi_privatisation_regions;
92 mc_model_checker->process().read_variable(
93 "smpi_privatisation_regions",
94 &remote_smpi_privatisation_regions, sizeof(remote_smpi_privatisation_regions));
95 s_smpi_privatisation_region_t privatisation_regions[process_count];
96 mc_model_checker->process().read_bytes(
97 &privatisation_regions, sizeof(privatisation_regions),
98 remote(remote_smpi_privatisation_regions));
100 std::vector<simgrid::mc::RegionSnapshot> data;
101 data.reserve(process_count);
102 for (size_t i = 0; i < process_count; i++) {
103 const simgrid::mc::RegionSnapshot* ref_privatized_region = nullptr;
104 if (ref_region && ref_region->storage_type() == StorageType::Privatized)
105 ref_privatized_region = &ref_region->privatized_data()[i];
106 data.push_back(simgrid::mc::region(region_type, start_addr,
107 privatisation_regions[i].address, size, ref_privatized_region));
110 simgrid::mc::RegionSnapshot region = simgrid::mc::RegionSnapshot(
111 region_type, start_addr, permanent_addr, size);
112 region.privatized_data(std::move(data));
113 return std::move(region);
118 void add_region(int index, simgrid::mc::Snapshot* snapshot,
119 simgrid::mc::RegionType type,
120 simgrid::mc::ObjectInformation* object_info,
121 void *start_addr, void* permanent_addr,
124 if (type == simgrid::mc::RegionType::Data)
125 xbt_assert(object_info, "Missing object info for object.");
126 else if (type == simgrid::mc::RegionType::Heap)
127 xbt_assert(!object_info, "Unexpected object info for heap region.");
129 simgrid::mc::RegionSnapshot const* ref_region = nullptr;
130 if (mc_model_checker->parent_snapshot_)
131 ref_region = mc_model_checker->parent_snapshot_->snapshot_regions[index].get();
133 simgrid::mc::RegionSnapshot region;
135 const bool privatization_aware = object_info
136 && mc_model_checker->process().privatized(*object_info);
137 if (privatization_aware && MC_smpi_process_count())
138 region = simgrid::mc::privatized_region(
139 type, start_addr, permanent_addr, size, ref_region);
142 region = simgrid::mc::region(type, start_addr, permanent_addr, size, ref_region);
144 region.object_info(object_info);
145 snapshot->snapshot_regions[index]
146 = std::unique_ptr<simgrid::mc::RegionSnapshot>(
147 new simgrid::mc::RegionSnapshot(std::move(region)));
151 static void get_memory_regions(simgrid::mc::Process* process, simgrid::mc::Snapshot* snapshot)
153 const size_t n = process->object_infos.size();
154 snapshot->snapshot_regions.resize(n + 1);
156 for (auto const& object_info : process->object_infos)
157 add_region(i++, snapshot, simgrid::mc::RegionType::Data,
159 object_info->start_rw, object_info->start_rw,
160 object_info->end_rw - object_info->start_rw);
162 xbt_mheap_t heap = process->get_heap();
163 void *start_heap = heap->base;
164 void *end_heap = heap->breakval;
166 add_region(n, snapshot, simgrid::mc::RegionType::Heap, nullptr,
167 start_heap, start_heap,
168 (char *) end_heap - (char *) start_heap);
169 snapshot->heap_bytes_used = mmalloc_get_bytes_used_remote(
171 process->get_malloc_info());
174 if (mc_model_checker->process().privatized() && MC_smpi_process_count())
175 // snapshot->privatization_index = smpi_loaded_page
176 mc_model_checker->process().read_variable(
177 "smpi_loaded_page", &snapshot->privatization_index,
178 sizeof(snapshot->privatization_index));
181 snapshot->privatization_index = simgrid::mc::ProcessIndexMissing;
184 #define PROT_RWX (PROT_READ | PROT_WRITE | PROT_EXEC)
185 #define PROT_RW (PROT_READ | PROT_WRITE)
186 #define PROT_RX (PROT_READ | PROT_EXEC)
188 /** \brief Fills the position of the segments (executable, read-only, read/write).
190 // TODO, use the ELF segment information for more robustness
191 void find_object_address(
192 std::vector<simgrid::xbt::VmMap> const& maps,
193 simgrid::mc::ObjectInformation* result)
195 char* name = xbt_basename(result->file_name.c_str());
197 for (size_t i = 0; i < maps.size(); ++i) {
198 simgrid::xbt::VmMap const& reg = maps[i];
199 if (maps[i].pathname.empty())
201 char* map_basename = xbt_basename(maps[i].pathname.c_str());
202 if (strcmp(name, map_basename) != 0) {
208 // This is the non-GNU_RELRO-part of the data segment:
209 if (reg.prot == PROT_RW) {
210 xbt_assert(!result->start_rw,
211 "Multiple read-write segments for %s, not supported",
212 maps[i].pathname.c_str());
213 result->start_rw = (char*) reg.start_addr;
214 result->end_rw = (char*) reg.end_addr;
216 // The next VMA might be end of the data segment:
217 if (i + 1 < maps.size()
218 && maps[i + 1].pathname.empty()
219 && maps[i + 1].prot == PROT_RW
220 && maps[i + 1].start_addr == reg.end_addr)
221 result->end_rw = (char*) maps[i + 1].end_addr;
224 // This is the text segment:
225 else if (reg.prot == PROT_RX) {
226 xbt_assert(!result->start_exec,
227 "Multiple executable segments for %s, not supported",
228 maps[i].pathname.c_str());
229 result->start_exec = (char*) reg.start_addr;
230 result->end_exec = (char*) reg.end_addr;
232 // The next VMA might be end of the data segment:
233 if (i + 1 < maps.size()
234 && maps[i + 1].pathname.empty()
235 && maps[i + 1].prot == PROT_RW
236 && maps[i + 1].start_addr == reg.end_addr) {
237 result->start_rw = (char*) maps[i + 1].start_addr;
238 result->end_rw = (char*) maps[i + 1].end_addr;
242 // This is the GNU_RELRO-part of the data segment:
243 else if (reg.prot == PROT_READ) {
244 xbt_assert(!result->start_ro,
245 "Multiple read only segments for %s, not supported",
246 maps[i].pathname.c_str());
247 result->start_ro = (char*) reg.start_addr;
248 result->end_ro = (char*) reg.end_addr;
252 result->start = result->start_rw;
253 if ((const void*) result->start_ro > result->start)
254 result->start = result->start_ro;
255 if ((const void*) result->start_exec > result->start)
256 result->start = result->start_exec;
258 result->end = result->end_rw;
259 if (result->end_ro && (const void*) result->end_ro < result->end)
260 result->end = result->end_ro;
261 if (result->end_exec && (const void*) result->end_exec > result->end)
262 result->end = result->end_exec;
264 xbt_assert(rresult->start_exec || result->start_rw || result->start_ro);
269 /************************************* Take Snapshot ************************************/
270 /****************************************************************************************/
272 /** \brief Checks whether the variable is in scope for a given IP.
274 * A variable may be defined only from a given value of IP.
276 * \param var Variable description
277 * \param frame Scope description
278 * \param ip Instruction pointer
279 * \return true if the variable is valid
281 static bool valid_variable(simgrid::mc::Variable* var,
282 simgrid::mc::Frame* scope,
285 // The variable is not yet valid:
286 if (scope->range.begin() + var->start_scope > (std::uint64_t) ip)
292 static void fill_local_variables_values(mc_stack_frame_t stack_frame,
293 simgrid::mc::Frame* scope,
295 std::vector<s_local_variable>& result)
297 simgrid::mc::Process* process = &mc_model_checker->process();
299 if (!scope->range.contain(stack_frame->ip))
302 for(simgrid::mc::Variable& current_variable :
305 if (!valid_variable(¤t_variable, scope, (void *) stack_frame->ip))
309 // FIXME, get rid of `region_type`
310 if ((long) stack_frame->ip > (long) process->libsimgrid_info->start_exec)
315 s_local_variable_t new_var;
316 new_var.subprogram = stack_frame->frame;
317 new_var.ip = stack_frame->ip;
318 new_var.name = current_variable.name;
319 new_var.type = current_variable.type;
320 new_var.region = region_type;
321 new_var.address = nullptr;
323 if (current_variable.address != nullptr)
324 new_var.address = current_variable.address;
325 else if (!current_variable.location_list.empty()) {
326 simgrid::dwarf::Location location =
327 simgrid::dwarf::resolve(
328 current_variable.location_list,
329 current_variable.object_info,
330 &(stack_frame->unw_cursor),
331 (void *) stack_frame->frame_base,
332 &mc_model_checker->process(), process_index);
334 if (!location.in_memory())
335 xbt_die("Cannot handle non-address variable");
336 new_var.address = location.address();
339 xbt_die("No address");
341 result.push_back(std::move(new_var));
344 // Recursive processing of nested scopes:
345 for(simgrid::mc::Frame& nested_scope : scope->scopes)
346 fill_local_variables_values(
347 stack_frame, &nested_scope, process_index, result);
350 static std::vector<s_local_variable> get_local_variables_values(
351 std::vector<s_mc_stack_frame_t>& stack_frames, int process_index)
353 std::vector<s_local_variable> variables;
354 for (s_mc_stack_frame_t& stack_frame : stack_frames)
355 fill_local_variables_values(&stack_frame, stack_frame.frame, process_index, variables);
356 return std::move(variables);
359 static std::vector<s_mc_stack_frame_t> unwind_stack_frames(simgrid::mc::UnwindContext* stack_context)
361 simgrid::mc::Process* process = &mc_model_checker->process();
362 std::vector<s_mc_stack_frame_t> result;
364 unw_cursor_t c = stack_context->cursor();
366 // TODO, check condition check (unw_init_local==0 means end of frame)
370 s_mc_stack_frame_t stack_frame;
372 stack_frame.unw_cursor = c;
376 unw_get_reg(&c, UNW_REG_IP, &ip);
377 unw_get_reg(&c, UNW_REG_SP, &sp);
382 // TODO, use real addresses in frame_t instead of fixing it here
384 simgrid::mc::Frame* frame = process->find_function(remote(ip));
385 stack_frame.frame = frame;
388 stack_frame.frame_name = frame->name;
389 stack_frame.frame_base =
390 (unw_word_t) frame->frame_base(c);
392 stack_frame.frame_base = 0;
393 stack_frame.frame_name = std::string();
396 result.push_back(std::move(stack_frame));
398 /* Stop before context switch with maestro */
399 if (frame != nullptr &&
400 frame->name == "smx_ctx_sysv_wrapper")
403 int ret = unw_step(&c);
405 xbt_die("Unexpected end of stack.");
407 xbt_die("Error while unwinding stack");
410 if (result.empty()) {
411 XBT_INFO("unw_init_local failed");
415 return std::move(result);
418 static std::vector<s_mc_snapshot_stack_t> take_snapshot_stacks(simgrid::mc::Snapshot* * snapshot)
420 std::vector<s_mc_snapshot_stack_t> res;
422 for (auto const& stack : mc_model_checker->process().stack_areas()) {
423 s_mc_snapshot_stack_t st;
425 // Read the context from remote process:
426 unw_context_t context;
427 mc_model_checker->process().read_bytes(
428 &context, sizeof(context), remote(stack.context));
430 st.context.initialize(&mc_model_checker->process(), &context);
432 st.stack_frames = unwind_stack_frames(&st.context);
433 st.local_variables = get_local_variables_values(st.stack_frames, stack.process_index);
434 st.process_index = stack.process_index;
436 unw_word_t sp = st.stack_frames[0].sp;
438 res.push_back(std::move(st));
441 (char*) stack.address + stack.size - (char*) sp;
442 (*snapshot)->stack_sizes.push_back(stack_size);
445 return std::move(res);
449 static void snapshot_handle_ignore(simgrid::mc::Snapshot* snapshot)
451 xbt_assert(snapshot->process());
454 for (auto const& region : mc_model_checker->process().ignored_regions()) {
455 s_mc_snapshot_ignored_data_t ignored_data;
456 ignored_data.start = (void*)region.addr;
457 ignored_data.data.resize(region.size);
458 // TODO, we should do this once per privatization segment:
459 snapshot->process()->read_bytes(
460 ignored_data.data.data(), region.size, remote(region.addr),
461 simgrid::mc::ProcessIndexDisabled);
462 snapshot->ignored_data.push_back(std::move(ignored_data));
466 for(auto const& region : mc_model_checker->process().ignored_regions())
467 snapshot->process()->clear_bytes(remote(region.addr), region.size);
471 static void snapshot_ignore_restore(simgrid::mc::Snapshot* snapshot)
473 for (auto const& ignored_data : snapshot->ignored_data)
474 snapshot->process()->write_bytes(
475 ignored_data.data.data(), ignored_data.data.size(),
476 remote(ignored_data.start));
479 static std::vector<s_fd_infos_t> get_current_fds(pid_t pid)
481 const size_t fd_dir_path_size = 20;
482 char fd_dir_path[fd_dir_path_size];
483 int res = snprintf(fd_dir_path, fd_dir_path_size,
484 "/proc/%lli/fd", (long long int) pid);
485 xbt_assert(res >= 0);
486 if ((size_t) res > fd_dir_path_size)
487 xbt_die("Unexpected buffer is too small for fd_dir_path");
489 DIR* fd_dir = opendir(fd_dir_path);
490 if (fd_dir == nullptr)
491 xbt_die("Cannot open directory '/proc/self/fd'\n");
493 std::vector<s_fd_infos_t> fds;
495 struct dirent* fd_number;
496 while ((fd_number = readdir(fd_dir))) {
498 int fd_value = xbt_str_parse_int(fd_number->d_name, "Found a non-numerical FD: %s. Freaking out!");
503 const size_t source_size = 25;
505 int res = snprintf(source, source_size, "/proc/%lli/fd/%s",
506 (long long int) pid, fd_number->d_name);
507 xbt_assert(res >= 0);
508 if ((size_t) res > source_size)
509 xbt_die("Unexpected buffer is too small for fd %s", fd_number->d_name);
511 const size_t link_size = 200;
513 res = readlink(source, link, link_size);
516 xbt_die("Could not read link for %s", source);
518 xbt_die("Buffer to small for link of %s", source);
523 if(smpi_is_privatisation_file(link))
527 // This is (probably) the DIR* we are reading:
528 // TODO, read all the file entries at once and close the DIR.*
529 if(strcmp(fd_dir_path, link) == 0)
532 // We don't handle them.
533 // It does not mean we should silently ignore them however.
534 if (strncmp(link, "pipe:", 5) == 0 || strncmp(link, "socket:", 7) == 0)
537 // If dot_output enabled, do not handle the corresponding file
538 if (dot_output != nullptr) {
539 char* link_basename = xbt_basename(link);
540 if (strcmp(link_basename, _sg_mc_dot_output_file) == 0) {
547 // This is probably a shared memory used by lttng-ust:
548 if(strncmp("/dev/shm/ust-shm-tmp-", link, 21)==0)
551 // Add an entry for this FD in the snapshot:
553 fd.filename = std::string(link);
554 fd.number = fd_value;
555 fd.flags = fcntl(fd_value, F_GETFL) | fcntl(fd_value, F_GETFD) ;
556 fd.current_position = lseek(fd_value, 0, SEEK_CUR);
557 fds.push_back(std::move(fd));
561 return std::move(fds);
564 simgrid::mc::Snapshot* take_snapshot(int num_state)
566 XBT_DEBUG("Taking snapshot %i", num_state);
568 simgrid::mc::Process* mc_process = &mc_model_checker->process();
570 simgrid::mc::Snapshot* snapshot = new simgrid::mc::Snapshot(mc_process);
572 snapshot->num_state = num_state;
574 for (auto& p : mc_model_checker->process().simix_processes())
575 snapshot->enabled_processes.insert(p.copy.pid);
577 snapshot_handle_ignore(snapshot);
579 if (_sg_mc_snapshot_fds)
580 snapshot->current_fds = get_current_fds(mc_model_checker->process().pid());
582 const bool use_soft_dirty = _sg_mc_sparse_checkpoint && _sg_mc_soft_dirty;
584 /* Save the std heap and the writable mapped pages of libsimgrid and binary */
585 get_memory_regions(mc_process, snapshot);
587 mc_process->reset_soft_dirty();
589 snapshot->to_ignore = mc_model_checker->process().ignored_heap();
591 if (_sg_mc_visited > 0 || strcmp(_sg_mc_property_file, "")) {
593 take_snapshot_stacks(&snapshot);
595 snapshot->hash = simgrid::mc::hash(*snapshot);
601 snapshot_ignore_restore(snapshot);
603 mc_model_checker->parent_snapshot_ = snapshot;
608 void restore_snapshot_regions(simgrid::mc::Snapshot* snapshot)
610 for(std::unique_ptr<s_mc_mem_region_t> const& region : snapshot->snapshot_regions) {
611 // For privatized, variables we decided it was not necessary to take the snapshot:
613 restore(region.get());
617 // TODO, send a message to implement this in the MCed process
618 if(snapshot->privatization_index >= 0) {
619 // Fix the privatization mmap:
620 s_mc_restore_message message;
621 message.type = MC_MESSAGE_RESTORE;
622 message.index = snapshot->privatization_index;
623 mc_model_checker->process().getChannel().send(message);
629 void restore_snapshot_fds(simgrid::mc::Snapshot* snapshot)
631 if (mc_mode == MC_MODE_SERVER)
632 xbt_die("FD snapshot not implemented in client/server mode.");
634 for (auto const& fd : snapshot->current_fds) {
636 int new_fd = open(fd.filename.c_str(), fd.flags);
638 xbt_die("Could not reopen the file %s fo restoring the file descriptor",
639 fd.filename.c_str());
640 if (new_fd != fd.number) {
641 dup2(new_fd, fd.number);
644 lseek(fd.number, fd.current_position, SEEK_SET);
648 void restore_snapshot(simgrid::mc::Snapshot* snapshot)
650 XBT_DEBUG("Restore snapshot %i", snapshot->num_state);
651 const bool use_soft_dirty = _sg_mc_sparse_checkpoint && _sg_mc_soft_dirty;
652 restore_snapshot_regions(snapshot);
653 if (_sg_mc_snapshot_fds)
654 restore_snapshot_fds(snapshot);
656 mc_model_checker->process().reset_soft_dirty();
657 snapshot_ignore_restore(snapshot);
658 mc_model_checker->process().clear_cache();
660 mc_model_checker->parent_snapshot_ = snapshot;